17 May 2022 in Computers & Security
5 November 2022 in Journal of Information Security and Applications
1 June 2021 in Physical Communication
29 September 2026 in SSRN Electronic Journal
The decoding accuracy of electroencephalography (EEG) signals is crucial to the performance of brain-computer interface (BCI) systems, yet it is severely bottlenecked by pervasive cross-distribution discrepancies in EEG signals. These discrepancies stem from two key sources—multi-subject/cross-session variability and intra-subject task-induced variability both eroding the generalization and robustness of decoding models. Although Euclidean-space alignment approaches offer an interpretable solution for reconciling such shifts, their integration into deep learning frameworks is constrained by two critical limitations: the lack of end-to-end trainability and adaptive adjustment mechanisms tailored to dynamic EEG characteristics. To address these gaps, this paper proposes a lightweight, model-agnostic Euclidean Space Alignment Net (ESAN) for cross-distribution EEG signal alignment, enabling task adaptability via flexible, end-to-end distribution matching. ESAN can be seamlessly embedded into mainstream EEG decoding backbones to optimize MI recognition and brainprint identification on the same dataset: it mitigates multi-subject EEG feature distribution shifts for MI recognition and intra-subject task-induced shifts for brainprint identification. Extensive experiments on three public MI-EEG datasets demonstrate that ESAN consistently improves the performance of diverse deep learning backbones: for multi-subject MI classification, the highest accuracy gains reach 4.01\%, 2.62\%, and 2.93\% on BCI IV-2a, BCI IV-2b, and High-Gamma, respectively; for cross-task brainprint identification, the highest accuracy improvements reach 10.28\%, 5.51\%, and 1.88\% on the three datasets, respectively, while also consistently outperforming representative alignment methods. This work enables efficient exploitation of cross-distribution EEG samples and advances the robustness and practical deployment of EEG-based BCI systems.
17 July 2026 in ACM Transactions on Privacy and Security
Federated Learning (FL) is a privacy-preserving distributed learning framework, but its distributed data collection and client-side training pipeline expose the global model to backdoor attacks. Existing FL backdoor attacks often depend on fixed triggers or on update manipulations that are easier to isolate under robust aggregation and model-inspection defenses. We propose CAEBA (Conditional AutoEncoder Backdoor Attack), a dynamic hidden backdoor framework that uses a conditional autoencoder to generate target-aware and visually stealthy triggers while progressively implanting the backdoor through federated optimization. CAEBA separates the clean classifier from the trigger generator, formulates the attack as a dual-constrained objective, and updates the generator at a lower frequency than the classifier to stabilize local optimization. We evaluate CAEBA on MNIST, FashionMNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet under representative aggregation rules and defenses, including FL-Detector, RFLBAT, DeepSight, FoolsGold, and FLAME. The results show that CAEBA preserves main-task accuracy while maintaining persistent backdoor effectiveness.